* Make the following private headers public:
* Lift/CPUStateAccessAnalysisPass.h
* Lift/CSVOffsets.h
* Lift/PTCDump.h
* Lift/VariableManager.h
* Move from revngSupport to revngLift:
* IRAnnotators.{h,cpp}
* SelfReferencingDbgAnnotationWriter.{h,cpp}
* Move from revngSupport to revngModel:
* FunctionTags.{h,cpp}
* ProgramCounterHandler.{h,cpp}
* Move from revngSupport to revngRecompile:
* OriginalAssemblyAnnotationWriter.{h,cpp}
This reduces the occasions where some LLVM optimization passes rewrite
the allocas, reducing the chances that the model type metadata is
discarded by LLVM passes that we don't control.
A pass that was doing this was InstCombine.
This commit further centralizes the creation of allocas + ptrtoint. The
creation of these pair of instructions is problematic when the size of a
pointer in the input and target architecture differ.
In fact, in such cases, LLVM turns trunc into masking the integer to
suppresse the high bits. This results in problematic code.
This commit introduces an assumption that ensures LLVM does not do that.
Ivan Krysak